MiniMax's new open M2.5 and M2.5 Lightning near state of the art while costing 1/20th of Claude Opus 4

Izzo Your AI bill just became irrelevant.

Izzo You're listening to Exploring Next, I'm Izzo, and this is episode one hundred eighty-four with Boone. And Boone, MiniMax just dropped their M2.5 model that delivers Claude Opus performance at one-twentieth the cost.

Boone I've been watching the benchmarks all morning, Izzo. This isn't just cheaper — it's matching Claude Opus 4.6 on SWE-Bench at eighty percent while costing fifteen cents per million input tokens versus Claude's five dollars.

Izzo That's the kind of math that changes everything overnight. We've been in this world where using frontier AI felt like hiring a brilliant but expensive consultant — you watch every token. Now suddenly you can run four AI agents continuously for a year for ten thousand dollars.

Boone And here's what gets me excited — MiniMax is already eating their own dog food. Thirty percent of all tasks at their company are handled by M2.5, and eighty percent of their committed code is generated by it.

Izzo *laughs* So they're literally building the model that's building itself. But let's dig into how they pulled this off technically, because this isn't just about throwing more compute at the problem.

Boone Right, it's all about their Mixture of Experts architecture. They've got 230 billion parameters total, but the clever part is they only activate 10 billion for any given token. So you get the reasoning depth of a massive model with the speed of something much smaller.

Izzo Okay, break that down for me — how does the model decide which experts to activate?

Boone Think of it like having a team of specialists. When you ask about Python code, it routes to the programming experts. Financial modeling? Different set of experts fire up. The routing network learns which combinations work best for different types of problems.

Izzo That's smart, but the real innovation seems to be in their training approach. They built this whole reinforcement learning framework called Forge specifically for this.

Boone Exactly — and this is where it gets interesting. Instead of just training on text, Forge creates thousands of simulated workspaces where the model actually practices coding, using tools, building real projects. It's learning by doing, not just by reading.

Izzo That explains why they're seeing such strong performance on agentic tasks. The model isn't just predicting the next token — it's learned to actually plan and execute work.

Boone They even developed this mathematical approach called CISPO — Clipping Importance Sampling Policy Optimization — to keep the model stable during all that intensive RL training. Without it, the model would overcorrect and become unstable.

Izzo And the results speak for themselves. Eighty percent on SWE-Bench, seventy-six percent on tool calling benchmarks. But what really matters is they're doing this while being open source under a modified MIT license.

Boone That licensing is clever — you can use it commercially but you have to display 'MiniMax M2.5' prominently in your UI. It's like open source with built-in marketing.

Izzo From a product perspective, this changes the entire playbook. Remember when we had to optimize every prompt to save costs? That constraint just evaporated. You can now throw high-reasoning models at routine tasks that were cost-prohibitive before.

Boone The speed improvements are huge too — they're seeing thirty-seven percent faster end-to-end task completion. That means agentic pipelines where models talk to other models finally move fast enough for real-time applications.

Izzo I'm giving this a solid A-minus. The only thing holding it back from an A-plus is we need to see how it performs in production at scale. But this pricing basically makes AI infrastructure a rounding error for most enterprises.

Boone What's really wild is seeing Chinese labs like MiniMax releasing models just days behind the US frontier. They're not just catching up — they're innovating on efficiency and cost in ways that might leapfrog the competition.

Izzo Alright, if this got your attention, here's what you should go build. First, grab the model from Hugging Face and run some local experiments — see how it handles your specific use cases.

Boone Second, if you're building agents, test their API with some real workflows. At fifteen cents per million tokens, you can afford to be experimental. I'm definitely adding an agent orchestration project to my weekend list.

Izzo And third, start thinking about all those tasks you've been doing manually because AI was too expensive. Document generation, code reviews, research synthesis — suddenly all of that becomes economically viable.

Boone The math just fundamentally changed, Izzo. We're moving from AI as an expensive specialist to AI as an affordable workforce.

Izzo When intelligence becomes too cheap to meter, everything changes. We'll be watching how this plays out in production. Until next time.